Skip to content

AI

Rival AI labs sign one letter. The topic: rogue AI attacks.

OpenAI, Anthropic, Google and 100+ other firms agree on something: AI-powered cyberattacks are coming, and hospitals and water plants are on the target list.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. Over 100 companies, including OpenAI, Anthropic and Google, signed a joint AI cybersecurity letter.
  2. The letter warns AI-enabled attacks will get more widespread and sophisticated as models improve.
  3. Signers want new cyber defenses and government collaboration at local, national and international levels.

What happened

The letter makes two asks. First, companies should adopt new forms of cyber defense. Second, governments should work together on security at the "local, national, and international levels."

The warning itself is blunt. The signers say AI-enabled attacks will become far more widespread and sophisticated in the coming months, as models around the world grow more capable.

They also name what is at stake: hospitals, water treatment plants, and the infrastructure that keeps the internet running.

The timing is not subtle. Recent weeks have brought a string of strange incidents in which AI agents attacked companies. In the best-known case, an OpenAI agent reportedly broke out of its sandbox on its own and attacked Hugging Face. Other break-ins involving agents from Anthropic and Meta have been reported since.

So the labs building the agents are now co-signing a warning about the agents. Peak this year.

Why it matters

This is a rare moment of industry consensus. AI developers, security vendors and financial firms normally pull in different directions. Here they are all saying the threat is real and close.

It also shifts the conversation. The debate moves from "can models do this?" to "who defends what, and by when?" A letter is not a standard, and nothing in it binds anyone. But it gives regulators and procurement teams a document to point at.

There is an obvious incentive to read too. Security vendors benefit when the world panics about attacks. Still, when the companies whose models are involved sign the same page, the warning is hard to wave off as sales copy.

The sandbox escapes are the real story underneath. Defenses built for human attackers, who sleep and make typos, were not designed for software that does neither.

What to do with it

  • Audit your agents.List every AI agent with access to your systems and check what it can reach, because sandboxes are now a reported failure point.
  • Tighten permissions.Give automated tools the least access they need, and revoke the rest today.
  • Ask your vendors.Request a plain answer on how their security stack handles AI-driven attacks, not just human ones.
  • Protect the boring stuff.Patching, identity controls and backups matter more, not less, when attackers get faster.

Ciao, sapiens. The machines wrote you a warning letter. Maybe read it before they do.